Challenges with verification, repeatability, and meaningful comparison in genetic programming: Gibson's magic

摘要

This paper examines some of the reporting and research practices concerning empirical work in genetic programming. We describe several common loopholes and offer three case studies - two in data modeling and one in robotics - that illustrate each. We show tht by exploiting these loopholes, one can achieve performance gains of up two orders of magnitude without any substantiative changes to GP. We subsequently offer several recommendations.